Squish Meets Structure — Designing With Shoggoth
summary
Elicit uses large language models to read PDFs, extract structured fields into spreadsheet rows and cut work that usually needs 10–20 people months; the system reports about 2–3% incorrect extractions by combining prompt engineering, fine‑tuning, RLHF, temperature control and compositional chains (breaking tasks into search → rank → extract → cite). Presenter cites GPT training mixes (≈15% books, ≈3% Wikipedia, ~97% English web crawl) and argues for soft shoggoth metaphors and small focused tools rather than one magic textbox.
AI-written summary · about
chapters (11)
- 0:00Introduction and context
- 2:31What is a language model?
- 5:34Capabilities and potential
- 9:17The case study of Elicit
- 13:07Failures and hallucinations
- 16:59The squishy nature of models
- 20:08Squish versus structure
- 23:28Design strategies and tools
- 26:54Compositionality
- 30:34Designing specific AI tools
- 34:10Q&A and closing thoughts
related talks
49mSquish Meets Structure – Designing with Shoggoth
26mWriting for humans in an AI world
40mRuby Writing Machines
43mMoldable Development & Glamorous Toolkit in Practice
19mSketchforme: Composing Sketched Scenes from Text Descriptions for Interactive Applications
16mWorking Against Predictability
30mUsing AI to Create Better Requirements and Designs
10mFrom Philosophy to Interfaces: an Explanatory Method and a Tool Inspired by Achinstein's Theory ...